Application-Specific Network-on-Chip Architecture
Customization via Long-Range Link Insertion
Umit Y. Ogras
Department of Electrical and Computer Engineering
Carnegie Mellon University
Pittsburgh, PA 15213-3890, USA
e-mail: uogras@ece.cmu.edu
Abstract Networks-on-Chip (NoCs) represent a promising
solution to complex on-chip communication problems. The
NoC communication architectures considered so far are based
on either completely regular or fully customized topologies. In
this paper, we present a methodology to automatically synthe-
size an architecture where a few application-specific long-range
links are inserted on top of a regular mesh network. This way,
we can better exploit the benefits of both complete regularity
and partial customization. Indeed, our experimental results
show that inserting application-specific long-range links signif-
icantly increases the critical traffic workload at which the net-
work state transits from a free to a congested regime. This, in
turn, results in a significant reduction in the average packet
latency and a major improvement in the achievable network
throughput.
I. INTRODUCTION
Continuous scaling of CMOS technology makes it possible
to put many heterogeneous devices on a single chip. Large-
scale integration of these blocks onto a single chip calls for
truly scalable Networks-on-Chip (NoC) communication archi-
tectures [1,2,3]. Regular NoC architectures based on grid-like
(or 2D lattice) topologies provide well-controlled electrical
parameters and reduced power consumption across the links.
However, due to the lack of fast paths between remotely situ-
ated nodes, such architectures may suffer from long packet
latencies. Indeed, having many hops between different commu-
nicating nodes, not only increases the message latency, but also
increases the message blocking probability thus making the
end-to-end packet latency more unpredictable. Consequently,
such generic platforms may become easily less attractive for
application-specific designs that need to guarantee a given level
of performance.
On the other hand, fully customized topologies [8-11] can
improve the overall network performance, but they distort the
regularity of the grid structure. This results in links with widely
varying lengths, performance and power consumption. Conse-
quently, better logical connectivity comes at the expense of a
penalty in the structured nature of the wiring which is anyway
one of the main advantages offered by the regular on-chip net-
works [1]. In the extreme case, fully customized solutions may
end up resembling ASIC-style designs where individual mod-
ules communicate by packet switching. Hence, the usual prob-
lems of cross-talk, timing closure, global wires etc. may
undermine the overall gain obtainable through customization.
Radu Marculescu
Department of Electrical and Computer Engineering
Carnegie Mellon University
Pittsburgh, PA 15213-3890, USA
e-mail: radum@ece.cmu.edu
Figure 1. Illustration of adding long-range links to a 4x4
standard mesh network.
Fortunately, these two extreme points in the design space
(i.e. designs based on purely regular or completely customized
topologies) are not the only solutions possible for NoC architec-
tures. In fact, it is interesting to note that many technological,
biological, and social networks are neither completely regular,
nor irregular [14,15]. One can view these networks as a super-
position of clustered nodes with short links and a collection of
random long-range links that produce “shortcuts” among differ-
ent regions of the network. Regular lattice networks with a
number of additional random long-range links, similar to the
one shown in Figure 1, can be used to model such networks
[15]. This paper explores precisely the potential of using stan-
dard mesh networks in conjunction with a few additional long-
range links to improve the performance of regular NoCs.
Inserting long-range links to regular architectures clearly
reduces the average distance between remote nodes. However,
inserting long-range links cannot be done randomly for NoCs
because adding extra links has a more pronounced, yet barely
studied, impact on the dynamic properties of the network,
which are characterized by traffic congestion. At low traffic
loads, the average packet latency exhibits a weak dependence
on the traffic injection rate. However, when the traffic injection
rate exceeds a critical value, the packet delivery times rise
abruptly and the network throughput starts collapsing
(Figure 2). The state of the network before congestion (i.e. the
region on left hand side of the critical value) is the free state,
while the state beyond the critical value (right hand side) is said
to be the congested state. Finally, the transition from the free
state to the congested one is known as phase transition region.
As it turns out, the phase transition in regular networks can
be significantly delayed by introducing additional long-range
links (see Figure 2) [16]. Due to the exponential increase in the
latency beyond the critical point, even a small right hand shift

Figure 2. Shift in the critical traffic load after the insertion of long-
range links to a 6x6 mesh network under hotspot traffic (Section V).
in the critical traffic value may result in orders of magnitude
reduction for latency. Similarly, the achievable throughput
grows with the right shift of the critical traffic value. This phe-
nomenon is at the very heart of our optimization technique.
The main objective of our optimization technique is to boost
the network performance (i.e. reduce the average packet latency
and increase the network throughput) by maximizing the value
of the critical traffic load via smart insertion of application-spe-
cific long-range links. To this end, our contribution is twofold:
• First, for a given application, we propose an algorithm that
determines the most beneficial long-range links to be inserted in
a mesh network.
• Second, we present a deadlock-free decentralized routing
algorithm that exploits the long-range links to achieve the
desired performance level.
The paper is organized as follows. Section II reviews
related work. The proposed approach for smart link insertion is
explained in Section III. Practical considerations in implement-
ing long-range links are discussed in Section IV, while the
experimental results appear in Section V. Finally, Section VI
concludes the paper by summarizing our main contribution.
II. RELATED WORK
The use of NoCs as a scalable communication architecture
is discussed in [1,2,3]. Design methodologies for application-
specific NoCs are discussed in [4-11]. Studies in [6,7] consider
regular network topologies and present algorithms for applica-
tion mapping under different routing strategies. On the other
hand, fully customized communication architecture synthesis
for a given application is addressed in [8,9,10].
To the best of our knowledge, the idea of optimizing a
generic grid-like network with application-specific long-range
links is first addressed in this paper. Previous work on a similar
idea comes from the networks theory side and uses very idealis-
tic assumptions. For instance, the authors of [16] investigate the
effect of adding random links to mesh and torus networks under
uniform traffic. The packets in the network consist of a single
atomic entity containing address information only. Also, due to
the infinite buffer assumption, the authors do not deal with
deadlock states explicitly.
In contrast to this prior work, we do consider wormhole
routing with arbitrary flit sizes and network routers with
bounded input buffers. Most importantly, instead of uniform
traffic, we assume application-specific traffic patterns and
present an algorithm which inserts the long-range links in a
smart manner rather than randomly. Due to the bounded input
buffers, additional long-range links may cause deadlock states.
For this reason, we also present a deadlock-free routing algo-
rithm that exploits the long-range links to achieve the desired
performance boost.
III. LONG-RANGE LINK INSERTION ALGORITHM
We start by formulating the long-range link inserting prob-
lem. After that, we present the details of the solution following
a top-down approach.
A. System model and basic assumptions
The system of interest consists of a set T of m × n tiles
interconnected by a 2D mesh network1. The tiles of the network
(referred to as PEs) are populated with processing and/or stor-
age elements that communicate with each other via the net-
work. We do not make any assumption about the distribution of
the packet injection rate into the network, but only consider the
frequencies at which the PEs communicate with each other.
Due to limited on-chip buffer resources and low latency
requirements, it makes sense to assume wormhole switching for
the network; the results derived here, however, are also applica-
ble to packet- and virtual cut-through switching networks. Fur-
ther, we do not assume any particular routing algorithm for the
mesh network; the only requirement is that the underlying rout-
ing algorithm has to be deadlock-free and minimal. Deadlock-
free property is desirable for on-chip networks for two reasons:
First, implementing deadlock detection and recovery mecha-
nisms is expensive in terms of silicon resources. Second, such
mechanisms can cause unpredictable delays, which need to be
avoided for most embedded applications.
After inserting the long-range links as explained in
Section C, the routers without extra links simply use the default
routing strategy. Since this default strategy cannot route the
packets across the newly added links, we define a deadlock-free
routing strategy which enables the use of the newly added long-
range links (Section E).
B. Problem formulation
The communication volume between the PE located at tile
i ∈ T and the PE located at tile j∈ T is denoted by Vij. We
compute the frequency of communication, fij, between PEs i
and j by normalizing the inter-tile communication volume as
follows:
1. The following discussion assumes a 2D mesh but the proposed tech-
nique is applicable to other topologies for which a distance definition as
in equations 6 and 7 (in Section III) exists.
 V ij
∀ i, j, p, q ∈ T f ij = --------------------------
∑ ∑ Vpq
p q≠p
The addition of long-range links introduces an overhead due
to the additional wires and repeaters1 connecting the wire seg-
ments to ensure the latency-insensitive operation. Hence, we
need to model the cost of long-range links to have a measure of
this overhead.
Without losing the generality, we measure the size of long-
range links, s ( l ) , in multiples of basic link units which are
identical to the regular links used in the mesh network. This is
reasonable, since the long-range links consist of a number of
standard links connected by repeaters, as shown in Figure 8.
The number of repeaters required by a long-range link is given
by s ( l ) – 1 . Consequently, the maximum amount of permissible
overhead can be expressed as a multiple of the standard link
segments that make up the long-range link. For example, a
resource constraint of S means that only long-range links con-
sisting of at most S units of standard links, total, can be added.
We can now state the application-specific long-range link
insertion problem as follows:
Given
• fij ∀i, j ∈ T
• Maximum number of links that can be added, S
• Routing strategy for the mesh network, R
Determine
• The set of long-range links to be added on top of the mesh
network, L S
• A deadlock-free routing strategy that governs the use of the
newly added long-range links, R L
such that
max ( λ c ) subject to ∑ s(l ) < S
l ∈ LS
where λ c is the critical load at which the network enters the
congested phase. To give some intuition, the long-range links
are added to maximize the critical traffic, λ c , subject to the
total amount of available resources; that is, the phase transition
region is delayed beyond the value a standard mesh network (of
exactly same size) can offer. Maximizing λ c increases the
achievable throughput and reduces the latency compared to the
original critical load, as shown later in the experiments.
Note that, the objective of inserting long-range links is by
no means limited to maximizing λ c . Indeed, other objective
functions (e.g. increased fault-tolerance, guaranteed service,
etc.), can replace or augment the objective of maximizing λc in
order to take the full advantage of the inserted links.
C. Iterative link addition algorithm
We propose an efficient iterative algorithm that inserts the
most beneficial long-range links to the current configuration of
the network, provided that the available resources are not used
up yet. The link insertion algorithm is summarized in Figure 3.
1. In our terminology, the repeaters act primarily as storage elements
(like FIFO buffers), as explained in Section IV in more detail.
(1)
Figure 3. The flow of the long-range link insertion algorithm. The
evaluation step is detailed in Section D.
The algorithm starts with a standard mesh network and
takes the communication frequencies between the network tiles,
the default routing algorithm and the amount of resources
allowed to use as inputs. Then, the algorithm selects all possible
pairs of tiles (i.e. C ( T , 2 ) pairs where T is the number of
nodes in the network) and inserts links between them. After
inserting each long-range link, the resulting network is evalu-
ated to find out the gain obtained over the previous configura-
tion. Since our goal is to maximize λ c , we compare different
configurations in terms of critical traffic load as detailed in
Section D. After the most beneficial long-range link is found,
the information about this link is stored and the amount of uti-
lized resources updated. For example, if a long-range link con-
sisting of four equivalent segments of standard links is added,
the utilization is incremented by four.
The procedure described above repeats until all available
resources are used up. Once this happens, an architecture con-
(2) figuration file is generated. Then, the routing strategy govern-
ing the use of long-range links is produced and written to a
routing configuration file, as described in Section E.
D. Evaluation of the critical traffic value
While the impact of routing strategy, switching techniques
and network topology on the critical point have been studied via
simulation [18], no work has been done to maximize the traffic
critical value subject to resource constraints. The major obstacle
in optimizing the critical load comes from the difficulty in mod-
elling the variation of critical point as a function of the design
decisions. Several theoreticians [16,17] estimate the criticality
point using mean field models. However, unlike our work, these
studies assume uniform traffic, infinite buffers, and the esti-
mates are valid only for regular grids without long-range con-
nections.
The key idea of our contribution is to reduce the estimation
of critical point of the network to just one parameter that can be
computed analytically, much faster than simulation. This is
important since using very accurate estimates obtained through
simulation would be simply too costly to use within an optimi-
zation loop. The optimization goal can still be achieved using
the simple parameter, as long as the comparison between two
network configurations matches the one with the critical load.
 In the following, we relate λ c to the free packet delay (τ0) 8x8 Mesh Network
in the network, which can be efficiently computed. Let the λ
number of messages in the network (at time t) be N ( t ) and the Nave / τave
aggregated packet injection rate be λ , i.e.

packets/cycle
Nave / τ0
λ = ∑ λi , λi :the injection rate of tile i ∈ T
i∈T
In the free state (i.e. when λ < λ c ), the network is in the
steady-state, so the packet injection rate equals the packet ejec-
tion rate. As a result, we can equate the injection and ejection
rates to obtain the following approximation:
N ave (3) λ (packets/cycle)
λ ≈ ----------- Figure 4. Experimental verification of Equation 3 and Equation 4
τ ave
for a 8x8 mesh network.
where τave is the average time each packet spends in the net- line with triangular markers in Figure 4 illustrates the upper
work, and Nave = <N ( t )> is the average number of packets in bound given in Equation 4. We observe that this expression pro-
the network. The exact value of τave is a function of the traffic vides a good approximation at lower data rates and holds the
injection rate, as well as topology, routing strategy, etc. While upper bound property.
no exact analytical model is available in the literature, we Computation of τ0
observe that τave shows a weak dependence on the traffic injec- For arbitrary traffic patterns characterized by the communi-
tion rate when the network is in the free state. Hence, τ0 can be cation frequencies fij ∀i, j ∈ T , τ0 can be written as
used to approximate τave. If we denote the average number of
c L
packets in the network, at the onset of the criticality, by Nave , τ0 = ∑ ∑ fij d ( i, j ) ( t r + t s + t w ) + max ( ts + t w ) -----
W
(5)
we can write the following relation: i j≠i
c
Nave (4) where d ( i, j ) is the distance from routers i to router j, and tr, ts,
λ c ≈ ------------- tw are the is architectural parameters representing time to make
τ0
the routing decision, traverse the switch and the link, respec-
This approximation is also an upper bound for the critical tively. Finally, L is the length of the packet, while W is the width
load λ c , since τ 0 ≤ τ ave ( λ c ). We note that λ = N ave ⁄ τ ave fol- of the network channel.
lows also Little’s law. Indeed, other theoretical studies proposed For the standard mesh network, the Manhattan distance
c
to approximate λ c = Nave ⁄ τ 0 using mean field [16] and distance (dM) is used to compute d ( i, j ), i.e.
models [17] under uniform traffic.
Given that the number of messages in the network, at the d M ( i, j ) = i x – j x + i y – j y (6)
c
onset of the criticality, is bounded by its capacity, Nave , the crit-
where subscripts x and y denote the x-y coordinates, respec-
ical traffic load λ c and the average packet latency are inversely
tively. For the routers with long-range links, an extended dis-
proportional to each other. Indeed, if the average packet latency
tance definition is needed in order to take the long-range
decreases, the phase transition point is delayed, as demonstrated
connections into account. Hence, we use the following general-
in Figure 2, where the reduction in the latency is obtained due
ized definition:
to the long-range links. Our optimization technique uses the
relationship between λ c and τave to maximize the critical load. ⎧ d M ( i, j ) if no long-range link is attached to i (7)
More specifically, we minimize τ0 which can be efficiently d ( i, j )= ⎨
⎩ min ( d M ( i, j ), 1 + d M ( k, j ) ) if l ( i, k ) exists
computed in the optimization loop, as opposed to λ c for which
there is no known analytical result to date. In this equation, l(i,k) means that node i is connected to
Experimental Verification of the Equations 3 and 4 node k via a long-range link. The applicability of the distance
For completeness, we verified experimentally both definition is illustrated in Figure 5. Note that the distance com-
Equation 3 and Equation 4, as shown in Figure 4. The dotted putation does not require any global knowledge thus making the
line shows the actual packet injection rate (λ), for reference. routing decision algorithm decentralized.
The solid line with the square marker is obtained for a 8 × 8
network under the hotspot traffic, as the ratio between the aver-
age number of packets in the network and the average packet
delay at that particular injection rate. From these plots, it can be
clearly seen that there is a good agreement between the actual
value obtained through simulation and the one predicted by
Equation 3 before entering the criticality. Since the network is
not in steady-state beyond the critical traffic value, the
Equation 3 does not hold for higher injection rates. As men-
tioned before, the exact value of the average packet delay at a
given load, τ ( λ ), can only be found by simulation. The dashed Figure 5. Illustration of distance definition (see Eqn. 7).
 Interestingly enough, Equation 4 also confirms that the
average number of hops between the nodes, a common perfor-
mance metric, has indeed a positive impact on the dynamic
behavior of the network.
E. Routing strategy for long-range links
The routers without any extra link use the default routing
strategy. Defining a strategy for the routers with long-range
links is a necessity dictated by two factors:
• Without a customized mechanism in place, the newly added
long-range links cannot be utilized at all by the default routing
strategy;
• Introducing long-range links may result in cycling depen-
dencies. Therefore, arbitrary use of these links may result in
deadlock states.
The routing strategy proposed in this section tries to pro-
duce minimal paths towards the destination by utilizing the
long-range links effectively, as shown in Figure 6. To this end,
we first check whether or not there exists a long-range connec-
tion to the current router. If there is one, the distance to the des-
tination with and without the long-range link is computed using
Equation 7. It is interesting to note that we can obtain global
improvements in the network dynamics by using local informa-
tion only.
If a newly added long-range link produces a shorter distance
to the destination, we check whether or not using this link may
cause deadlock before accepting it as a viable route. In order to
guarantee that using the newly added link does not cause a
deadlock state, some limitations on the use of long-range links
are introduced.
We achieve deadlock-free operation by extending the turn-
model [13] to long-range links. More precisely, in the original
turn model, one out of four possible turns is prohibited to avoid
cyclic dependencies (Figure 7). However, unlike standard links,
the long-range links can extend in two directions, such as NE,
NW, etc. For this reason, one has to consider the rotation from
middle directions, NE, NW, SE, SW to the main directions N,
S, E and W. In the extended model, we arbitrarily chose the S-
to-E, S-to-W, SE-to-E, SE-to-W, SW-to-E, and SW-to-W turns
Figure 6. The description of the routing strategy.
as being illegal, as in Figure 7. Therefore, we check whether or
not the long-range links cause any of these prohibited turns. If it
is legal to use a long-range link, then the packet is forwarded to
this link. Otherwise, the default routing algorithm is employed.
Figure 7. Possible routing directions, basic and extended turn models.
A long-range link may become a traffic attractor and jam
the network earlier than the regular short links. For this reason,
if the routing algorithm is not adaptive (as it is our case), one
more check point is needed before assigning a traffic stream to
a new link. For instance, by assessing the amount of traffic
assigned to a long-range link, further traffic can be routed over
the link only if it is not likely to become a bottleneck.
IV. PRACTICAL CONSIDERATIONS
We analyze next the implications of customizing the regular
network architecture with long-range links on the actual design
implementation.
A. Implementation of long-range links
In order to preserve the advantages of structured wiring, the
long-range links are segmented into regular, fixed-length, net-
work links connected by repeaters. The use of repeaters with
buffering capabilities guarantees latency-insensitive operation,
as discussed in [19]. Repeaters can be thought as simplified
routers consisting of only two ports that accept an incoming flit,
stores it in a FIFO buffer, and finally forwards it to the output
port, as illustrated in Figure 8. If the depth of buffers in the
repeaters is at least 2 flits, then the packets can be effectively
pipelined to take the full advantage of the long-range links [20].
The final consideration in terms of implementation is the
increased size of the routers with extra links due to increased
number of ports (Figure 8). To measure the area overhead, we
implemented routers with 5 and 6 ports using Verilog and syn-
thesized them for a 1M gate Xilinx Virtex2 FPGA. The router
with 5 ports utilizes 387 slices (about 7% of total resources),
while the one with 6 ports utilizes 471 slices of the target
device. We also synthesized a pure 4×4 mesh network, and a
4×4 mesh network with 4 long-range links and observed that
the extra links induce about 10% area overhead. This overhead
has to be taken into account, while computing the maximum
number of long-range links that can be added to the regular
mesh network. While there is no theoretical limitation imposed
by our approach on the number of additional links a router can
have, a maximum of one long-range link per router is used in
our experiments. This way the regularity is minimally altered,
while still providing significant improvements over the stan-
dard mesh networks, as explained in Section V.

(a)
(b)
Figure 8. Implementation of the repeaters. Routers 1 and 3 are both
connected by Router 2 (a), the underlying mesh network, and the
inserted long-range link (b).
B. Energy-related considerations
One can measure the energy consumption using the Ebit
metric [12], defined as the energy required to transmit one bit of
information from the source to the destination. Ebit is given by
E bit = E L + E B + E (8)
bit bit S bit
where E Lbit , E Bbit and E S bit represent the energy consumed by
the interconnect, buffering and switching in the router, respec-
tively. Analyzing the energy consumption before and after the
insertion of long-range links shows that the proposed approach
does not induce a significant penalty in the total communication
energy consumption of the network. Indeed, since the long-
range links consist of several regular links with repeaters
between them (instead of routers), the link energy consumption
stays approximately the same whether the traffic flows over the
long-range link or over the original path provided by the mesh
network. On the other hand, the energy consumption due to the
switch and routing logic is greatly simplified in the repeater
design compared to the original routers. This results in a reduc-
tion in the switching energy consumption. Finally, the routers
with extra links will have slightly increased energy consump-
tion due to the larger crossbar switch.
We compared the energy consumption obtained by simula-
tion before and after the insertion of the long-range links, for
the traffic patterns reported in Section 6. We observed that the
link and buffer energy consumptions increase about 2% after
the insertion of long-range links, while the switch energy con-
sumption drops about 7%, on average. The results show that the
overall energy consumption increases by only about 1%.
The worst-case complexity of the technique, that is link
α
insertion and routing table generation, is O ( SN ) where
2< α < 3 . The run-time of the algorithm for the examples ana-
lyzed ranges from 0.14sec for a 4×4 network to less than half
hour for a 8 × 8 network on a Pentium III machine with 768MB
memory under Linux OS.
A. Experiments with synthetic traffic workloads
We first demonstrate the effectiveness of adding long-range
links to standard mesh networks by using synthetic traffic
inputs. Table 1 compares the critical traffic load, average packet
latency and throughput at the edge of criticality under hotspot
traffic pattern for 4×4 and 6 × 6 networks. For the hotspot traf-
fic three nodes are selected arbitrarily to act as hotspot nodes.
Each node in the network sends packets to these hotspot nodes
with a higher probability compared to the remaining nodes.
Critical load Latency at the critical load
(packet/cycle) (cycles)
λ Mc λ Lc L M ( λ Mc ) L L ( λ Mc )
hotspot 4x4 0.41 0.50 196.9 34.4
hotspot 6x6 0.62 0.75 224.5 38.2
Table 1: Critical load (packet/cycle) and latency comparison
(cycles) for regular mesh (M) and mesh with long links (L).
As shown in Table 1, inserting 4 long-range links (consist-
ing of 10 short link segments) to a 4×4 network makes the
phase transition region shift from 0.41 packet/cycle to 0.50
packet/cycle (the resulting network appears in Figure 1). Simi-
larly, the average packet latency at 0.41 packet/cycle injection
rate drops from 196.9 to 34.4 cycles. We also show the variation
of the network throughput and average packet latency as a func-
tion of traffic injection rate using a much denser scale in
Figure 9.
Hotspot Benchmark

V. EXPERIMENTAL RESULTS
We present next an extensive experimental study involving
a set of benchmarks with synthetic and real traffic patterns. The
NoCs under study are simulated using an in-house cycle accu-
rate C++-based NoC simulator developed specifically for this
project. The simulator models the long-range links precisely as
explained in Section IV. The configuration files describing the
additional long-range links and the routing strategy for a given
traffic pattern are generated using the proposed technique and Figure 9. Traffic injection rate vs. average packet latency and
network throughput for hotspot traffic is shown. The
supplied to the simulator as an input. improvement in terms of critical point and latency values at
criticality are indicated on the plots.
 Similar results have been obtained for a 6 × 6 grid, as
shown in Table 1. In this case, the phase transition region shifts
from 0.62 packet/cycle to 0.75 packet/cycle. Likewise, with the
addition of long-range links, the average packet latency at 0.62
packet/cycle injection rate drops from 224.5 to 38.2 cycles.
Comparison with the torus network
We also compared the performance of the proposed
approach against that achievable with a torus network, which
provides wrap around links added in a systematic manner. Our
simulations show that application-specific insertion of only 4
long-range links, with an overhead of 12 extra standard link
segments, provides 4% improvement in the critical traffic load
compared to a 4×4 torus under hotspot traffic. Furthermore,
the average packet latency at the critical point of the torus net-
work, 0.48 packet/cycle, drops from 77.0 to 34.4 cycles. This
significant gain is obtained over the standard torus network by
utilizing only half of the additional links, since the extra links
are inserted in a smart way considering the underlying applica-
tion rather than blindly adding wrap-around channels.
Scalability Analysis
To evaluate the scalability of the proposed technique, we
also performed experiments involving networks of sizes rang-
ing from 4×4 to 10 ×10 . Figure 10 shows that the proposed
technique results in consistent improvements when the network
size scales up. For example, by inserting only 6 long-range
links, consisting of 32 regular links in total, the critical load of a
10 ×10 network under hotspot traffic shifts from 1.18 packet/
cycle to 1.40 packet/cycle giving a 18.7% improvement. This
result is similar to the gain obtained for smaller networks.
Figure 10(a) also reveals that the critical traffic load grows with
the network size due to the increase in the total bandwidth.
Likewise, we observe consistent reduction in the average packet
latency across different network sizes, as shown in
Figure 10(b).
B. Experiments involving real applications
In this section, we evaluate the performance of the link
insertion algorithm using two applications with realistic traffic:
A 4x4 auto industry benchmark and a 5x5 telecom benchmark
retrieved from E3S benchmark suite [21].
The variation of average packet latency and network
throughput as a function of traffic injection rates for auto indus-
try benchmark is given in Figure 11. These plots show that the
insertion of long-range links shifts the critical traffic load from
0.29 packet/cycle to 0.33 packet/cycle resulting in a 13.6%
improvement. Similarly, as shown in Table 2, we observe that
the average packet latency for the network with long-range
links is consistently lower compared to that of a pure mesh net-
work. For instance, at 0.29 packet/cycle injection rate, the
latency drops from 98.0 cycles to 30.3 cycles giving about
69.0% reduction.
Similar improvements have been observed for the telecom
benchmark as shown in Figure 12. Specifically, the critical traf-
fic load is delayed from 0.44 packet/cycle to 0.60 packet/cycle
showing a 36.3% improvement due to the addition of long-
range links. Likewise, the latency at 0.44 packet/cycle traffic
injection rate drops from 73.1 cycles to 28.2 cycles (Table 2).
(a) (b)
Figure 10. Scalability results. Performance of the proposed technique
for larger network sizes.
Comparison with the mesh network with extra buffers
Implementation of long-range links requires buffers in the
repeaters. To demonstrate that the savings are the result of using
the long-range links, we also added extra amount of buffers to
the corresponding channels of the pure mesh network, equal to
the amount of buffers utilized for the long-range links.
Table 2 summarizes the results for standard mesh network
(M), standard mesh network with extra buffers (MB) and the
network with long-range links (L). We observe that insertion of
buffers improves the critical load by 3.5% for the auto industry
benchmark. On the other hand, the corresponding improvement
due to long-range links is 13.6% over initial mesh network and
10% over the mesh network with additional buffers. Likewise,
we note that with the insertion of long-range links, the average
packet latency reduces by 69% compared to the original value
and 57.0% compared to the mesh network with extra buffers.
Consistent results have been obtained for the synthetic traf-
fic workloads mentioned in the previous section and for the
telecom benchmark. Due to the limited space, we report only
the results for telecom benchmark (Table 2). The results show
that, with the addition of extra buffers, the critical traffic point
shifts only from 0.44 packet/cycle to 0.46 packet/cycle. Insert-
ing long-range links, on the other hand, shifts the critical point
to 0.60 packet/cycle which is a huge improvement in the net-
work capability. Similarly, the average packet latency obtained
by the proposed technique is almost 1/3 of the latency provided
by standard mesh and about 1/2 of the latency provided by
mesh with extra buffers.
Critical load Latency at critical load
(packet/cycle) (cycles)
L ( λ Mc )
λc
auto-indust M 0.29 98.0
auto-indust MB 0.30 70.5
auto-indust L 0.33 30.3
telecom M 0.44 73.1
telecom MB 0.46 56.0
telecom L 0.60 28.2
Table 2: Critical load (packet/cycle) and latency (cycles)
comparison for pure mesh (M), mesh with extra buffers
(MB) and mesh with long links (L).
 Figure 11. Traffic rate vs. packet latency and network
throughput for auto industry benchmark.
VI. CONCLUSION AND FUTURE WORK
We have presented a design methodology to insert applica-
tion-specific long-range links to standard grid-like networks. It
is analytically and experimentally demonstrated that insertion
of long-range links has an important impact on the dynamics, as
well as static properties of the network. Specifically, additional
long-range links increase the critical traffic workload. We have
also demonstrated that this increase means significant reduction
in the average packet latency in the network, as well as
improvement in the achievable throughput.
Our current work employs oblivious routing to utilize the
long-range links. We plan to extend this work to employ adap-
tive routing instead. Other possible extensions include inserting
long-range links for different objective functions such as fault-
tolerance and QoS operation.
Acknowledgements: This research is supported by
Marco GSRC, NSF CCR-00-93104, and SRC 2004-HJ-1189.
VII. REFERENCES
[1] W. Dally, B. Towles, “Route packets, not wires: On-chip intercon-
nection networks,” In Proc. DAC, June 2001.
[2] L. Benini, G. De Micheli. “Networks on chips: A new SoC para-
digm,” IEEE Computer. 35(1), 2002.
[3] A. Jantsch, H. Tenhunen (Eds.). Networks on Chip. Kluwer,
2003.
[4] M. Millberg, E. Nilsson, R. Thid, S. Kumar, A. Jantsch, “The
Nostrum backbone - a communication protocol stack for networks on
chip,” In Proc. VLSI Design, Jan. 2004.
[5] K. Goossens, et. al. “A Design flow for application-specific net-
works on chip with guaranteed performance to accelerate SOC
design and verification,” In Proc. DATE, March 2005.
[6] J. Hu, R. Marculescu, “Exploiting the routing flexibility for
energy/performance aware mapping of regular NoC architectures,”
In Proc. DATE, March 2003.
Figure 12. Traffic injection rate vs. average packet
latency and network throughput for telecom benchmark.
[7] S. Murali, G. De Micheli, “SUNMAP: A tool for automatic topol-
ogy selection and generation for NoCs,” In Proc. DAC, June 2004.
[8] A. Pinto, L. P. Carloni, A. L. Sangiovanni-Vincentelli, “Efficient
synthesis of networks on chip,” In Proc. ICCD, Oct. 2003.
[9] K. Srinivasan, K. S. Chatha, G. Konjevod “Linear programming
based techniques for synthesis of Network-on-Chip architectures,” In
Proc. ICCD, Oct. 2004.
[10] U. Y. Ogras, R. Marculescu, “Energy- and performance- driven
customized architecture synthesis using a decomposition approach,”
In Proc. DATE, March 2005.
[11] A. Jalabert, S. Murali, L. Benini, G. De Micheli, “xpipesCom-
piler: A tool for instantiating application specific networks on chip,”
In Proc. DATE, March 2004.
[12] T. T. Ye, L. Benini, G. De Micheli, “Analysis of power consump-
tion on switch fabrics in network routers,” In Proc. DAC, June 2003.
[13] C. J. Glass, L. M. Ni, “The turn model for adaptive routing,” In
Proc. ISCA, May 1992.
[14] D. J. Watts, S. H. Strogatz, “Collective dynamics of 'small-
world' networks,” Nature, 393:440-442, 1998.
[15] J. Kleinberg, “The Small-World phenomenon and decentralized
search,” SIAM News 37(3), April 2004.
[16] H. Fuks, A. Lawniczak, “Performance of data networks with
random links,” Mathematics and Computers in Simulation, vol 51,
1999.
[17] M. Woolf, D. Arrowsmith, R. J. Mondragon, J. M. Pitts, “Opti-
mization and phase transitions in a chaotic model of data traffic,”
Physical Review E, vol. 66 (2002).
[18] J. Duato, S. Yalamanchili, N. Lionel, Interconnection Networks:
an Engineering Approach. Morgan Kaufmann, 2002.
[19] L. P. Carloni, K. L. McMillan, A. L. Sangiovanni-Vincentelli,
“Theory of latency-insensitive design,” IEEE Trans. on CAD of Inte-
grated Circuits and Systems. 20(9), 2001.
[20] V. Chandra, A. Xu, H. Schmit, L. Pileggi, “An interconnect
channel design methodology for high performance integrated cir-
cuits,” In Proc. DATE, Feb. 2004.
[21] R. Dick, “Embedded system synthesis benchmarks suites
(E3S),” http://helsinki.ee.princeton.edu/dickrp/e3s/.